Mistral just introduced Mistral Large 4, a 1-trillion parameter natively multimodal open-weight model with 52 billion active parameters. Nicknamed "le Chonk," it scores 82% on real-world vulnerability reproduction tests and rivals frontier closed systems in visual grounding. Why it matters: it delivers sovereign, high-performance AI for cybersecurity and enterprise workflows, with weights dropping later this month.
Researchers at Zhejiang University just unveiled AudioHijack, a technique that embeds imperceptible malicious commands into everyday audio to hijack generative voice assistants. By slightly tweaking sound waveforms, attacks achieve up to a 96% success rate in forcing models to leak data or run unauthorized tasks, transferring even to commercial services from Microsoft and Mistral.
Mistral just rolled out Magistral Small 1.2 and Medium 1.2, giving its reasoning models visual capabilities to analyze uploaded images. The 24B parameter Small version can even be quantized to run locally on a 32GB MacBook or a single RTX 4090, offering strong reasoning without the cloud.
Mistral just rolled out a major update to its Le Chat chatbot, adding a deep research mode, multilingual reasoning, and advanced image editing. The real differentiator: it connects to enterprise on-premises data directly, letting businesses run deep analysis without shipping sensitive files to the cloud.
Mistral just launched Mistral Code, an AI coding assistant built on the open-source Continue project to take on Cursor and GitHub Copilot. It bundles in-house models like Codestral and Devstral, with a private beta now rolling out for VS Code and JetBrains development platforms. The big draw for enterprise teams is deployment flexibility: companies can run it in the cloud, on private servers, or entirely air-gapped.
Mistral just unveiled Devstral, a 24B coding model built with All Hands AI under a fully commercial Apache 2.0 license. It fits on a single RTX 4090 or a 32GB Mac, beating Google's Gemma 3 27B and DeepSeek's V3 on SWE-Bench Verified. Why it matters: Developers finally get a capable, agent-ready code model they can run locally without restrictive licenses.
Mistral just rolled out Mistral Medium 3, an efficiency-focused model claiming to match at least 90% of Claude 3.7 Sonnet's benchmark performance at $0.40 per million input tokens and $2 per million output tokens. Alongside the release, Mistral also made its corporate assistant, Le Chat Enterprise, generally available.
Mistral just introduced Mistral Saba, a 24-billion-parameter model tailored for Arabic-speaking countries. According to internal tests, it handles Arabic far better than Mistral Small 3 and also performs well with Indian languages like Tamil. It is accessible through Mistral's API and supports on-premise deployment.
Mistral just partnered with global news agency AFP to plug up-to-date journalism directly into its Le Chat assistant. The multi-year deal gives the bot access to AFP's archives dating back to 1983 alongside 2,300 daily stories across six languages, excluding photos and videos. It marks Mistral's first media pact as it pushes beyond raw models to take on ChatGPT.
Sakana AI just unveiled Transformer², an adaptive system that adjusts LLM weights at test time to tackle different tasks like a living brain. Using Singular Value Finetuning and RL, it beats LoRA across math and coding benchmarks while requiring far fewer parameters.
ElevenLabs just teamed up with Mistral, Perplexity, Hugging Face, and over a dozen other platforms to launch the AI Engineer Pack. Builders get $50 or more in credits across each participating tool, taking the financial sting out of testing multiple paid APIs when prototyping new projects.
Mistral just rolled out a new content moderation API powered by a fine-tuned Ministral 8B model. It classifies raw or conversational text across nine safety categories in multiple languages, covering everything from hate speech to PII. The startup also launched a batch API that cuts query costs by 25%.
